An adaptive graph convolutional network with residual attention for emotion recognition

Electroencephalogram (EEG)-based emotion recognition is a reliable and deployable method for identifying human emotional states. Currently, Graph convolution networks (GCN) have exhibited superior performance in extracting topological features of EEG. However, how to capture the dynamic topological relationship is still a challenge. In this paper, we propose an adaptive GCN with residual attention (AGC-RSTA) to extract the spatio-temporal discriminative features. Firstly, we construct an adaptive adjacency matrix in graph convolution, extracting the dynamic spatial topological features. We then utilize the residual spatio-temporal attention module to capture deep spatio-temporal features. Ablation studies and comparative experiments on the SEED and SEED-IV datasets demonstrate that our proposed model outperforms state-of-the-art methods, achieving recognition accuracies of 94.91% and 91.17%, respectively.